acceptodds
Under review as a conference paper at ICLR 2027

Farsighted Correction of Shortcut Learning: From Dynamic Starvation to Confidence-Gap Optimization

Abstract

Deep neural networks frequently rely on easily accessible spurious correlations, driving a phenomenon known as shortcut learning. Remarkably, this reliance remains invisible on standard loss curves, which decrease smoothly as models build brittle representations. In this work, we uncover the mathematical root of this invisibility by analyzing optimization dynamics via the empirical Neural Tangent Kernel (eNTK) along the training trajectory. We establish the non-asymptotic **Dynamic Starvation Theorem**, proving that a dominant shortcut eigenvalue restricts causal parameter updates to a bounded regime in the worst case of unbounded shortcut dominance. We further characterize how this starvation mechanism manifests under different loss functions: under **squared loss**, joint and ablated training yield identical predictions in output space, trapping the causal subnetwork in a degenerate null space as loss vanishes; under **cross-entropy loss**, self-correction is merely logarithmic. To resolve these failure modes, we propose **Farsighted Signal Optimization** (FSO), which reinjects counterfactual causal gradients into the training objective. We extend this to model-agnostic settings via GAP-FSO, proving that inter-group confidence gaps serve as exact proxies for shortcut parameters without explicit architectural splits. Unlike prior theoretical treatments of shortcut learning, which validate their predictions on simple architectures, we test GAP-FSO on ResNet-50 and DistilBERT across six vision and language benchmarks. Results are competitive with established baselines overall and clearly outperform them on the NLP benchmark.

Primary area:
learning theory
Keywords:
Empirical Neural Tangent Kernel, eNTK, shortcut learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_K4FBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how shortcuts affect a model's ability to learn genuinely useful features. The authors first propose FSO for models with separate causal and shortcut branches, providing additional training signals to the causal branch. They then introduce GAP-FSO for standard neural networks, which suppresses shortcuts by penalizing the difference in logit margins between aligned and conflicting samples

Critical strengths
  1. The paper analyzes the training dynamics under squared loss and cross-entropy loss separately, showing that the shortcut problems under these two losses are not the same.

  2. Beyond synthetic experiments, the authors also compare their method with group-robust approaches using ResNet, DistilBERT, and multiple vision and text datasets.

Critical weaknesses
  1. GAP-FSO interprets the difference in average logit margins between the two groups as a measure of shortcut strength. However, this equality relies on the assumption that the two groups have the same causal contribution and that the shortcut contribution changes sign exactly. In real-world datasets, even when samples are grouped by class, the difficulty of samples in the two groups may differ. In this case, reducing the gap does not necessarily mean removing the shortcut; it may also suppress useful predictive signals.

  2. The connection between the theory and the practical method is not fully explained. FSO promotes learning in the causal branch through additional gradients, while GAP-FSO, which is actually used for ResNet and DistilBERT, penalizes the logit gap between groups. According to the paper's own derivation under cross-entropy loss, the latter does not change the updates to the causal term. Therefore, the accuracy improvements on real-world datasets do not necessarily demonstrate that GAP-FSO alleviates the causal starvation described in the paper.

  3. The current experiments mainly report final accuracy, with limited validation of the underlying mechanism. The paper does not directly show whether reducing the gap also reduces the model's reliance on shortcuts such as color and background. Comparisons with other inter-group constraints under the same experimental settings are also missing.

Questions for authors

Could the authors conduct an experiment where the main object in an image is kept fixed while only the color or background is changed, and compare how the model's predictions change before and after training? It would also be useful to report whether these prediction changes correlate with changes in the logit gap. This would provide a more direct test of whether GAP-FSO achieves the claimed shortcut correction than simply comparing final accuracy.

5 comments
AuthorsOpenReview ↗

Response to W1 (1/2): Confidence Gap under Difficulty Asymmetry

A: We thank the reviewer for this insightful question. The equality Δ(t)=2uf(t)\Delta(t)=2u_f(t) indeed holds only under the theoretical idealizations of our framework (Proposition 5.1), where the causal contributions of the two groups are equal and cancel out under Assumption 2.1 (A1). On real-world data, this generalizes to Δ=2uf+δc\Delta=2u_f+\delta_c, where δc\delta_c denotes the difficulty gap between the groups, and the confidence gap alone may not disentangle the two terms. Similar idealizations are common in the theoretical literature on shortcut learning [1–4], so we examine empirically how a difficulty gap affects our method. Regarding class-level difficulty, the class-conditional (J=2J=2) form compares aligned and conflicting samples within each class, so differences in difficulty between classes cancel and only the difficulty within a class remains. We therefore conduct controlled experiments on within-class difficulty asymmetry, examining both downstream performance and internal representations.

To this end, we construct a controlled Dogs-vs-Cats benchmark by applying Gaussian blur with different probabilities to the aligned and conflicting groups (none, moderate, and strong asymmetry, i.e., blur with probabilities (qa,qc)=(0.5,0.5)(q_a,q_c)=(0.5,0.5), (0.25,0.75)(0.25,0.75) and (0.1,0.9)(0.1,0.9) on the aligned and conflicting images), making the conflicting group progressively harder. On test data, the resulting difficulty gap δc\delta_c, estimated via a shortcut-free reference model of the same architecture, accounts for roughly 1%, 14%, and 25% of the ERM gap under CE loss. As shown in Table 1, GAP-FSO consistently outperforms ERM in worst-group accuracy across all difficulty levels, demonstrating its effectiveness even under confounding difficulty asymmetries. Although the margin narrows as asymmetry grows because ERM itself improves under stronger blur, we hypothesize that because blur correlates with group identity, models can exploit difficulty itself as an additional predictive cue (e.g., relying on background darkness only when the image is sharp), aiding performance on the harder samples.

Table 1: Worst-group accuracy on Dogs-vs-Cats under varying difficulty asymmetry (CE loss).

Difficulty asymmetryERMGAP-FSOGain
None (symmetric blur)0.773 ± 0.0220.926 ± 0.002+0.154
Moderate0.824 ± 0.0130.947 ± 0.004+0.122
Strong0.886 ± 0.0080.963 ± 0.006+0.077
AuthorsOpenReview ↗

Response to W1 (2/2): Counterfactual Sensitivity and Core Probes

To examine the reliance on the shortcut independently of difficulty, we use a counterfactual swap: we fix the core subject (e.g., the animal or core text content) and alter only the surface shortcut attribute (e.g., rendering the image dark or bright, or substituting identity terms). Because the core content is identical, sample difficulty is held fixed, and any resulting shift in the model's output can be attributed solely to the shortcut. Based on this swap, we introduce two metrics using the decision margin zz (logit difference for CE; scalar output for MSE) and label y∈−1,+1y\in\\{-1,+1\\}: the average signed margin change cf_gap =E[y(zaligned−zconflicting)]=\mathbb{E}[y(z_{\text{aligned}}-z_{\text{conflicting}})], and the average absolute score change cf_abs =E∣zaligned−zconflicting∣=\mathbb{E}|z_{\text{aligned}}-z_{\text{conflicting}}| (used for text). Values closer to 0 indicate weaker shortcut reliance, where 0 means no change in the model's output; unlike the non-negative cf_abs, cf_gap is signed and therefore captures both the magnitude and direction of shortcut reliance. Since both metrics inherit the model's output scale, we compare ERM and GAP-FSO within the same model rather than across different backbones.

Turning to the empirical results in Table 2, GAP-FSO drives Δ\Delta toward zero while cf_gap shifts from 0.262 without asymmetry to −0.055 under moderate and −0.450 under strong asymmetry, whereas ERM remains high between 3.1 and 3.9. This negative shift reflects a mild over-correction of the original shortcut consistent with Δ=2uf+δc\Delta=2u_f+\delta_c: when the conflicting group is intrinsically harder (δc>0\delta_c > 0), forcing Δ→0\Delta \to 0 implies uf<0u_f < 0. Consequently, when within-class difficulty is tied to group identity, the confidence gap alone cannot be interpreted strictly as shortcut reliance. Even so, GAP-FSO's worst-group accuracy continues to improve under stronger asymmetry, achieving even higher performance than when cf_gap is positive (Table 1). This is consistent with our earlier hypothesis that the induced difficulty asymmetry itself acts as an informative predictive signal, suggesting that the model may leverage sample difficulty to aid classification on harder groups despite the apparent over-correction.

Table 2: cf_gap on Dogs-vs-Cats under varying difficulty asymmetry (CE loss).

Difficulty asymmetryERMGAP-FSO
None (symmetric blur)3.854 ± 0.5360.262 ± 0.046
Moderate3.134 ± 0.260−0.055 ± 0.046
Strong3.342 ± 0.486−0.450 ± 0.123

In addition, to verify whether the reduction in shortcut reliance comes at the cost of core features, we fit linear probes on the frozen penultimate features alongside counterfactual sensitivity evaluations (cf_gap, cf_abs). Table 3 summarizes the results across both datasets, where the counterfactual swaps demonstrate that GAP-FSO substantially reduces the model's reliance on the shortcut. Meanwhile, the probes address a different question (whether this reduction compromises core information) and indicate that it does not: the core probe, defined as the worst-group test accuracy of a linear read-out from the frozen features, remains at least as high as that of ERM. Therefore, GAP-FSO successfully mitigates shortcut reliance without compromising core representation quality, confirming that the performance gains stem from genuine robust feature learning rather than collateral damage to core task information.

Table 3: Evaluation of ERM and GAP-FSO on Dogs-vs-Cats and CivilComments (CE loss).

Dataset (model)MethodWorst-group acc.cf_gapcf_absCore probe
Dogs-vs-Cats (ResNet-50)ERM0.967 ± 0.0021.427 ± 0.0961.866 ± 0.0700.978 ± 0.004
Dogs-vs-Cats (ResNet-50)GAP-FSO0.989 ± 0.0010.165 ± 0.0230.263 ± 0.0230.989 ± 0.001
CivilComments (DistilBERT)ERM0.457 ± 0.048–1.942 ± 0.0880.682 ± 0.030
CivilComments (DistilBERT)GAP-FSO0.687 ± 0.062–0.641 ± 0.2490.691 ± 0.004
AuthorsOpenReview ↗

Response to W2 (1/2): How GAP-FSO Connects to Causal Starvation

A: We believe that, under the experimental design of these benchmarks, evaluating the model on out-of-distribution (worst-group) data is itself a partial verification of what the model has learned: since the shortcut no longer correlates with the label at test time, a higher accuracy indicates that the model relies more on core features rather than on spurious ones. The effect of GAP-FSO on causal starvation is indirect: in the idealized model, the penalty drives the shortcut component to zero (uf→0u_f\to0) instead of leaving the permanent offset of uncorrected CE, so that the shortcut no longer absorbs the shared error signal. Under this premise, a higher OOD accuracy accompanied by a lower reliance on the shortcut is consistent with GAP-FSO alleviating causal starvation indirectly.

While OOD accuracy provides an indirect signal, we also conducted counterfactual evaluations, which measure this reliance directly. At fixed difficulty, cf_gap and cf_abs are much smaller than those of ERM on every model family that we evaluated (Table 4): cf_gap falls from 1.427 to 0.165 on Dogs-vs-Cats (ResNet-50) and from 3.657 to 0.764 on Waterbirds, cf_abs falls from 1.942 to 0.641 on CivilComments (DistilBERT), CLIP ViT-B/32 and Qwen2.5-0.5B show the same direction, and on Colored-MNIST cf_gap falls from 8.033 to 0.661. In Table 5, the share of examples whose predicted label flips when only the shortcut is swapped (flip rate) falls as well on the image tasks, for example from 11.9% to 4.4% on Waterbirds (ResNet-50), and changes from 13.2% to 12.4% on CivilComments with DistilBERT. More results can be seen in the table.

Table 4: Sensitivity of the prediction to the shortcut under the counterfactual swap (CE loss).

Model and datasetSwapMetricERMGAP-FSO
CNN, Colored-MNISTcolour of the digitcf_gap8.033 ± 0.7890.661 ± 0.177
ResNet-50, Dogs-vs-Catsdark / brightcf_gap1.427 ± 0.0960.165 ± 0.023
ResNet-50, Waterbirdswater / landcf_gap3.657 ± 0.1170.764 ± 0.078
DistilBERT, CivilCommentsidentity termcf_abs1.942 ± 0.0880.641 ± 0.249
CLIP ViT-B/32, Dogs-vs-Catsdark / brightcf_gap3.132 ± 0.1930.145 ± 0.202
CLIP ViT-B/32, Waterbirdswater / landcf_gap5.485 ± 1.5700.339 ± 0.235
Qwen2.5-0.5B, CivilCommentsidentity termcf_abs3.583 ± 1.1450.186 ± 0.102

Table 5: Flip rate under the counterfactual swap (share of test examples whose predicted label changes when only the shortcut is swapped; CE loss).

Model and datasetSwapERMGAP-FSO
ResNet-50, Dogs-vs-Catsdark / bright0.028 ± 0.0010.005 ± 0.001
ResNet-50, Waterbirdswater / land0.119 ± 0.0140.044 ± 0.009
DistilBERT, CivilCommentsidentity term0.132 ± 0.0090.124 ± 0.017
CLIP ViT-B/32, Dogs-vs-Catsdark / bright0.047 ± 0.0070.023 ± 0.007
CLIP ViT-B/32, Waterbirdswater / land0.224 ± 0.0460.069 ± 0.009
Qwen2.5-0.5B, CivilCommentsidentity term0.096 ± 0.0150.072 ± 0.012
Reviewer_BnF1OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper analyzes shortcut learning using the training dynamics of the empirical Neural Tangent Kernel (eNTK), showing how dominant shortcut features can suppress learning of core features. It proposes FSO to restore counterfactual learning signals and GAP-FSO to reduce shortcut reliance through a group confidence-gap penalty. Experiments across six synthetic, vision, and language benchmarks evaluate the proposed mechanisms and robustness improvements.

Critical strengths

The time-varying eNTK perspective provides an interpretable account of feature competition. Connecting this analysis to a simple group confidence-gap penalty is appealing, and evaluation on both vision and language tasks strengthens its practical relevance.

Critical weaknesses

Corollary C.1 claims identical empirical-risk trajectories for joint and ablated training, but its proof establishes only identical limiting risk. Equal limiting predictions do not imply equal finite-time loss trajectories, leaving a central theoretical claim unsupported.

The empirical evaluation does not sufficiently establish the method’s added value over recent work. The baselines largely predate 2024, omitting closely related methods [1]. Moreover, the evaluation relies on modified benchmark protocols and a single NLP task, with limited stress testing of group-label noise and minority-group scarcity. These gaps leave the method’s comparative benefits and scope of applicability unclear.

  1. Spurious Correlation-Aware Embedding Regularization for Worst-Group Robustness
Questions for authors

How does Corollary C.1 establish equality of the entire risk trajectories rather than only their limits? If this equality does not hold, which conclusions about loss-curve indistinguishability remain valid?

When aligned and conflicting groups differ in core-feature difficulty, how can the confidence gap distinguish shortcut reliance from these differences? What evidence validates this interpretation in the ResNet-50 and DistilBERT experiments?

8 comments
AuthorsOpenReview ↗

Response to W2 (1/3): Comparison with SCER

A: We thank the reviewer for this constructive feedback. We agree that comparing against recent work is essential, and we address all raised points below, including a comparison with SCER [1], a second NLP task, benchmark protocol details, and stress tests on group-label noise and minority-group scarcity.

Before presenting the results, we hope to clarify how our starting point differs from existing debiasing literature. Rather than proposing an engineering-driven, state-of-the-art debiasing algorithm, our primary contribution is a theoretical account of how shortcuts form and suppress causal learning through training dynamics. GAP-FSO serves as the theoretically grounded intervention demonstrating that targeting this mechanism measurably reduces shortcut reliance, so we treat baselines as reference points rather than competitors in a benchmark race. Nevertheless, our subsequent empirical results demonstrate that our method remains highly competitive and exhibits robust performance across larger-scale models. Our analysis further suggests a broader design principle for future methods, which involves designing interventions by explicitly targeting training dynamics and gradient-level signal competition (such as error signal starvation), rather than relying on static geometric assumptions or heuristic group reweighting.

(a) Comparison with SCER. While the reviewer highlights SCER as a closely related baseline, its formulation differs substantially from ours in both theoretical motivation and analytical space. Specifically, GAP-FSO is derived from an eNTK and training-dynamics perspective, analyzing learning behavior in the output space and directly penalizing the confidence gap between aligned and conflicting data groups. By contrast, SCER operates statically in the embedding space under Gaussian mixture assumptions, measuring the shift induced by spurious attributes along the classifier direction. Furthermore, inspecting the official SCER source code reveals several unablated engineering modifications that are not documented in their paper, such as taking element-wise absolute group-mean differences scaled to unit norm and adding an identity shift to the covariance matrix. These heuristic adjustments alter both the direction and the metric of the penalty to boost empirical performance. (Even so, we evaluated SCER directly using its official source code without modifying its objective or heuristics in our experiments.) In contrast, GAP-FSO adheres strictly to its theoretical derivation without introducing auxiliary engineering tricks or explicit worst-group reweighting, simply adding a single signed penalty term with one weight γ\gamma to standard CE or MSE loss to evaluate the standalone value of intervening in the Dynamic Starvation Theorem mechanism. When evaluated under the exact same protocol (using the identical optimizer, model architecture, random seeds, and dataset splits), GAP-FSO performs on par with SCER across benchmarks (Table 1). These results demonstrate that GAP-FSO remains considerably competitive even without auxiliary worst-group reweighting or unablated implementation heuristics. We have cited and discussed SCER in the revised manuscript.

Table 1: Test worst-group accuracy, CE loss, mean ± std (ResNet-50 for Waterbirds and Dogs-vs-Cats, DistilBERT for CivilComments).

MethodWaterbirdsDogs-vs-CatsCivilComments
ERM0.784 ± 0.0150.967 ± 0.0020.457 ± 0.048
IRM0.843 ± 0.0130.988 ± 0.0020.648 ± 0.056
V-REx0.840 ± 0.0110.989 ± 0.0010.651 ± 0.036
GroupDRO0.853 ± 0.0070.989 ± 0.0010.694 ± 0.026
DFR0.863 ± 0.0120.976 ± 0.0010.659 ± 0.040
GAP-FSO (J=1J=1)0.835 ± 0.0270.989 ± 0.0010.687 ± 0.062
GAP-FSO (J=2J=2)0.843 ± 0.0070.987 ± 0.0020.711 ± 0.051
SCER0.829 ± 0.0080.986 ± 0.0020.662 ± 0.028
AuthorsOpenReview ↗

Response to W2 (2/3): A Second NLP Task, Larger Models, and Standard CelebA

(b) A second NLP task and larger models. We evaluate on the MultiNLI dataset using DistilBERT and Qwen2.5-0.5B backbones, and additionally on CivilComments with Qwen2.5-0.5B and on Dogs-vs-Cats and Waterbirds with CLIP ViT-B/32 (Table 2). GAP-FSO improves over ERM in every one of the 10 (dataset, loss) combinations, on every seed (by roughly 0.03 on Dogs-vs-Cats, where ERM is already strong, and by 0.13 to 0.34 elsewhere), and is on par with or above GroupDRO except on Qwen-MultiNLI under MSE, which we conjecture relates to Qwen's large-scale pre-training that is known to confer greater robustness to spurious correlations [2]. Compared to SCER, GAP-FSO has the higher mean worst-group accuracy under CE in all five cells, but the margin is within seed noise on Dogs-vs-Cats (0.968 vs. 0.965) and Qwen-MultiNLI (0.656 vs. 0.654); it is larger on DistilBERT-MultiNLI (0.518 vs. 0.461) and Qwen-CivilComments (0.739 vs. 0.611). SCER is inapplicable to MSE as its public implementation only supports CE loss. Across all transformer experiments, GAP-FSO has a higher mean than GroupDRO in 9 of 10 (dataset, loss) combinations, so our dynamics-driven penalty remains competitive when scaled to larger pretrained models.

Table 2: Test worst-group accuracy, mean ± std over 3 seeds.

BackboneDatasetLossERMGAP-FSOGroupDROSCER
CLIP ViT-B/32Dogs-vs-CatsCE0.936 ± 0.0210.968 ± 0.0080.953 ± 0.0080.965 ± 0.006
CLIP ViT-B/32Dogs-vs-CatsMSE0.938 ± 0.0160.964 ± 0.0050.955 ± 0.004–
CLIP ViT-B/32WaterbirdsCE0.704 ± 0.0100.848 ± 0.0070.832 ± 0.0510.827 ± 0.051
CLIP ViT-B/32WaterbirdsMSE0.714 ± 0.0410.841 ± 0.0370.811 ± 0.019–
DistilBERTMultiNLICE0.207 ± 0.0260.518 ± 0.0440.481 ± 0.0460.461 ± 0.005
DistilBERTMultiNLIMSE0.193 ± 0.0090.531 ± 0.0470.449 ± 0.066–
Qwen2.5-0.5BCivilCommentsCE0.476 ± 0.0240.739 ± 0.0450.682 ± 0.0970.611 ± 0.110
Qwen2.5-0.5BCivilCommentsMSE0.452 ± 0.0530.695 ± 0.0180.627 ± 0.016–
Qwen2.5-0.5BMultiNLICE0.436 ± 0.0050.656 ± 0.1080.630 ± 0.0230.654 ± 0.055
Qwen2.5-0.5BMultiNLIMSE0.402 ± 0.0560.632 ± 0.0380.710 ± 0.027–

(c) On the standard CelebA setting. We use Blond Hair × Heavy Makeup because with Male, (blond, male) is only 0.85% of the training split and the aligned/conflicting indicator has its y=1y=1 branch backwards (lines 1694–1697; all line numbers in our responses refer to the revised manuscript). The manuscript also explains that when a shortcut correlates with the label in opposite directions across classes the pooled gap can cancel, which the class-conditional (J=2J=2) form avoids (lines 450–454, 1581–1584); on our split J=2J=2 gains +8.2 and +7.8 points (lines 1602–1605, Table 4 in the manuscript). On the standard Blond Hair × Male protocol the new experiment agrees (Table 3 below), where the class-conditional form reaches 0.744 (CE) and 0.774 (MSE) against 0.457 and 0.448 for ERM, while the pooled form, which the theory predicts to be ill-suited here, stays at the ERM level. The class-conditional result is comparable to IRM (0.765, CE) and above GroupDRO (0.456), while DFR reaches 0.847 (the highest). We have added this experiment to the revised manuscript.

Table 3: Standard CelebA (Blond Hair × Male), test worst-group accuracy, mean ± std over 3 seeds.

MethodCEMSE
ERM0.457 ± 0.0250.448 ± 0.023
GroupDRO0.456 ± 0.0330.470 ± 0.042
DFR0.847 ± 0.0090.843 ± 0.002
IRM0.765 ± 0.0450.567 ± 0.017
V-REx0.733 ± 0.0710.430 ± 0.061
GAP-FSO (J=1J=1, pooled)0.426 ± 0.0280.448 ± 0.028
GAP-FSO (J=2J=2, class-conditional)0.744 ± 0.0350.774 ± 0.003

References

[2] Tu, L., Lalwani, G., Gella, S., & He, H. (2020). An empirical study on robustness to spurious correlations using pre-trained language models. Transactions of the Association for Computational Linguistics, 8, 621–633.

AuthorsOpenReview ↗

Response to W2 (3/3): Group-Label Noise and Minority-Group Scarcity

(d) Group-label noise and minority-group scarcity. We first conducted a label-noise experiment where the aligned or conflicting label of each training sample is flipped with probability ρ\rho, while the validation and test sets remain clean (Table 4). Using GroupDRO, which also relies on group labels, as a reference, one generally expects performance to degrade as ρ\rho increases. However, this trend does not strictly hold across all settings, as on Waterbirds GAP-FSO achieves slightly higher worst-group accuracy at ρ=0.2\rho=0.2 than at ρ=0.1\rho=0.1 (0.816 vs. 0.801). We will conduct further experiments across additional seeds and noise levels to verify whether a consistent monotonic trend exists, and we will include these findings in the relevant discussion (Appendix D.3, lines 1617–1637).

Table 4: Test worst-group accuracy vs. group-label flip rate ρ\rho (CE loss).

DatasetMethodρ=0\rho=0ρ=0.1\rho=0.1ρ=0.2\rho=0.2
CivilCommentsGAP-FSO0.687 ± 0.0620.584 ± 0.0480.438 ± 0.032
CivilCommentsGroupDRO0.694 ± 0.0260.581 ± 0.0160.519 ± 0.016
WaterbirdsGAP-FSO0.835 ± 0.0270.801 ± 0.0040.816 ± 0.007
WaterbirdsGroupDRO0.853 ± 0.0070.828 ± 0.0100.814 ± 0.000

Next, to evaluate whether GAP-FSO remains effective when minority groups are extremely scarce, we performed a second experiment that retains all aligned training samples while subsampling conflicting samples to a fraction f=10f=10\\%. As shown in Table 5, GAP-FSO performs comparably to GroupDRO, outperforming it by 3.6% on Waterbirds. We also observe that all methods degrade severely on CivilComments under this setting, likely because the model simply memorizes the tiny pool of remaining conflicting examples rather than learning generalizable causal features. We have added these results and discussion to the revised appendix.

Table 5: Minority-group scarcity experiment, test worst-group accuracy (CE loss).

DatasetERMGAP-FSOGroupDRO
Waterbirds0.567 ± 0.0180.661 ± 0.0040.625 ± 0.012
CivilComments0.170 ± 0.0110.200 ± 0.0000.231 ± 0.027
Reviewer_xdvTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper tackles the problem of classification with spurious correlations and shortcut learning. The main observation is that when shortcut features are much easier to learn than the actual features, they quickly reduce most of the training error, leaving very little signal for the model to learn the actual features. The authors call this dynamic starvation and theoretically analyze this behavior under different loss functions. To resolve the problem, they first propose FSO, which tries to preserve the learning signal that the actual features would receive if the shortcut was not present. Since separating shortcut and actual features inside a real network is difficult, they then propose a practical version called GAP-FSO, which uses known groups of shortcut-aligned and shortcut-conflicting samples and reduces the difference in model confidence between them. They evaluate the approach on several classification datasets, including Colored MNIST, Waterbirds, CelebA, Dogs-vs-Cats, and CivilComments, and show competitive results compared with existing approaches.

Critical strengths

The paper tackles an important and still relevant problem of shortcut learning and spurious correlations. I find the observation about the training loss particularly interesting: when shortcut features are easier to learn, they can quickly reduce the loss and consequently leave very little learning signal for the actual features. While this observation is intuitive, the paper provides a theoretical formulation and analysis that makes the phenomenon more precise. I also appreciate that the proposed method is directly motivated by this observation, rather than being an independent empirical solution.

Critical weaknesses

The main limitation is that the practical method requires prior information about which samples are associated with the spurious or shortcut features. If this information is already available, it raises the question of whether the shortcut can be addressed more directly using existing approaches. Another limitation is that the evaluation mainly focuses on conventional classification settings and benchmarks. Given the increasing importance of foundation models, LLMs, and multimodal models, it would be valuable to investigate whether the proposed observations and methods remain applicable when shortcuts are more complex and difficult to identify.

Another concern is that the paper uses substantial jargon and mathematical notation to explain concepts that often have relatively simple underlying intuitions. This makes parts of the paper unnecessarily difficult to follow, particularly for readers from different backgrounds. I recommend introducing the intuition behind the main concepts first, before presenting the corresponding mathematical formalism.

Questions for authors

Since GAP-FSO requires knowing which samples are aligned or conflicting with the spurious feature, what is the advantage over directly balancing, reweighting, or removing the spurious signal when possible like by masking or weakening the spurious features?

How can the proposed method be used when we do not know the spurious feature beforehand and do not have group annotations? What if there are multiple spurious features and annotating one leaves the other still in the data?

Do you expect the same observation and method to work for modern foundation models, LLMs, or vision-language models, where the spurious features can be much more complex and difficult to identify?

Can you provide experiments on more modern foundation models to show that the proposed method goes beyond conventional classification benchmarks such as Waterbirds and Colored MNIST?

8 comments
AuthorsOpenReview ↗

Response to W1: Reliance on Shortcut Knowledge and Applicability to Foundation Models

A: We thank the reviewer for raising these two limitations. Since they coincide with the reviewer's questions, we address them in detail in our answers below, and kindly refer the reviewer there. The reliance on prior knowledge of the shortcut is discussed in our answers to Q1 (the advantage over directly balancing, reweighting, or masking the spurious signal) and Q2 (the setting without group annotations, including a new label-free experiment, and the case of multiple shortcuts). The applicability to foundation models, LLMs, and multimodal models is discussed in our joint answer to Q3 and Q4, where we add new experiments on CLIP ViT-B/32 and Qwen2.5-0.5B.

AuthorsOpenReview ↗

Response to W2: Intuitions Added Before the Formalism

A: We thank the reviewer for this helpful suggestion and agree that several parts of the paper present the formalism before the intuition behind it. In the revised manuscript, we have opened each main formal result with a short plain-language preview of what it shows and why it matters (marked in blue; all line numbers in our responses refer to the revised manuscript). Specifically, we added the following intuitions:

  • Subnetwork competition (Section 2.2, Assumption 2.1, lines 140–143). We view the network as two pathways whose outputs are added together: a shortcut pathway (e.g., reading the background) and a causal pathway (e.g., reading the bird's shape). Both learn from the same remaining error, and the constants σf,σc\sigma_f,\sigma_c simply describe how easy each pathway is to learn.
  • Starvation bound (Section 3.1, Theorem 3.1, lines 176–179). Both pathways learn only from the error that is still left. If the easy shortcut pathway removes most of this error early, little is left for the causal pathway, so its total learning is capped no matter how long training continues. The Rayleigh activity is simply how fast each pathway uses up the remaining error.
  • Loss-dependent regimes (Propositions 3.2 and 3.3, lines 250–253). Under squared loss, learning stops once the training error reaches zero, so the causal pathway stays under-trained permanently. Under CE loss, the loss never reaches exactly zero, so the causal pathway keeps learning, but its learning signal shrinks as the model becomes confident, which makes the recovery extremely slow.
  • FSO and its null-space geometry (Section 4.1, lines 318–324). When the training error is zero, there are infinitely many ways to split the credit for a prediction between the two pathways, and the loss cannot tell them apart. Because the shortcut is always learned much faster, standard training without any prevention easily ends up relying on it unintentionally, settling on a split in which the shortcut takes most of the credit. FSO adds the requirement that the causal pathway alone must also fit the data, which leaves only one valid split: the one in which the causal pathway does all the work.
  • Confidence gap (Section 5.1, Proposition 5.1, lines 401–404). We moved the existing intuition before the proposition: the causal pathway gives the same vote to aligned and conflicting examples, whereas the shortcut's vote flips sign between them, so subtracting the two group averages cancels the causal part and leaves twice the shortcut part.
  • GAP-FSO under CE (Proposition 5.2, lines 418–422). Under CE, standard training keeps a permanent shortcut offset because the shortcut is correct on most training examples, and the CE gradient that could correct it vanishes as the model becomes confident. The GAP-FSO penalty pulls this offset back to zero with a force that depends only on how far the offset is from zero, not on the model's confidence, so it keeps acting after the CE gradient has vanished.

We also simplified the terminology: for example, "candidate shortcut subnetwork" and "causal complement" are now simply the "shortcut subnetwork" and the "causal subnetwork", each defined in plain words (Section 2.2, lines 145–146), consistent with their use in the rest of the paper. Similarly, we now explain the "trajectory-realized eNTK" at its first use as the tangent kernel measured along the actual training path rather than frozen at initialization (Introduction, lines 51–53), gloss "persistent excitation" in Theorem 3.1 as both subnetworks continuing to consume the error at some minimum rate (lines 204–205), and replace "fiber" and "harmful/harmless redundancy" in Section 4.1 and Figure 3 with a plain description of which freedoms change the model's predictions and which do not (lines 282–289 and 345–352). We believe these previews will make the paper accessible to readers from different backgrounds while keeping the formal results intact.

AuthorsOpenReview ↗

Response to Q1: Advantage over Balancing, Reweighting, and Masking

A: We thank the reviewer for this insightful question. While balancing, reweighting, and masking are natural and effective remedies at the data or input level [1–3], GAP-FSO offers a distinct scientific and methodological contribution by intervening directly in the optimization dynamics. First, whereas existing approaches modify the data distribution or mask out spurious features, GAP-FSO is derived directly from the optimization failure mechanism identified by our theory, namely that the shortcut subnetwork absorbs the shared error signal early in training and starves the causal subnetwork (Dynamic Starvation Theorem). FSO counteracts this by injecting causal gradients, and GAP-FSO realizes this without network splitting by penalizing the confidence gap between aligned and conflicting groups, a quantity that equals the shortcut contribution in our idealized model (Proposition 5.1). Second, prior theoretical studies on shortcut learning typically rely on idealized settings, such as linear models, random features, or small networks [4–6], much like idealized models in physics that isolate core mechanisms before addressing real-world complexity. Importantly, these analyses, like ours, must specify in advance which feature is the shortcut and which is causal, because the competition between the two cannot even be defined without designating them. Knowing the shortcut (equivalently, the aligned/conflicting partition) is therefore a prerequisite of the theoretical derivation rather than an extra requirement introduced by GAP-FSO, and our method inherits exactly the information that the theory assumes. Following this scientific tradition, our primary goal is to rigorously understand why shortcuts form and suppress causal learning under squared and CE losses, rather than to propose a state-of-the-art debiasing algorithm. However, going beyond standard theoretical validations on toy models, we additionally evaluate GAP-FSO across substantially more complex architectures, including ResNet-50, DistilBERT, CLIP ViT-B/32, and Qwen2.5-0.5B (with the latter models added during the rebuttal), to examine how our theoretically derived intervention behaves and measurably reduces shortcut reliance in modern deep networks. Finally, beyond this specific regularizer, our analysis suggests a broader design principle for future methods, which involves shifting from modifications of the data or the loss weighting toward training-time monitoring and intervention in the competition for the shared error signal.

References

[1] Kirichenko, P., Izmailov, P., & Wilson, A. G. (2022). Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937.

[2] Idrissi, B. Y., Arjovsky, M., Pezeshki, M., & Lopez-Paz, D. (2022). Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning (pp. 336–351). PMLR.

[3] Sagawa, S., Koh, P. W., Hashimoto, T. B., & Liang, P. (2020). Distributionally robust neural networks. In International Conference on Learning Representations.

[4] Pezeshki, M., Kaba, O., Bengio, Y., Courville, A. C., Precup, D., & Lajoie, G. (2021). Gradient starvation: A learning proclivity in neural networks. Advances in Neural Information Processing Systems, 34, 1256–1272.

[5] Sagawa, S., Raghunathan, A., Koh, P. W., & Liang, P. (2020). An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning (pp. 8346–8356). PMLR.

[6] LaBonte, T., & Muthukumar, V. (2026). SGD Provably Prioritizes a Shortcut Spurious Feature in the XOR Model. arXiv preprint arXiv:2606.30444.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to Q1: Counterfactual Swap and Its Correlation with the Gap

A: We have discussed the requested experiment, the counterfactual swap, in our answers to W1 and W2. We read "before and after training" as without and with the penalty (ERM against GAP-FSO) on the same data, architecture, and seeds. We hold the subject fixed and change only the shortcut: the color of the digit on Colored-MNIST, dark or bright images on Dogs-vs-Cats, the same segmented bird on a water or land background on Waterbirds, and identity-term substitution (e.g., black and white, male and female) on CivilComments. With the metrics cf_gap and cf_abs defined in our answer to W1, the sensitivity of the prediction to the shortcut falls from ERM to GAP-FSO on every model family that we evaluated (Table 4), and so does the share of predictions that flip when only the shortcut is swapped on the image tasks (Table 5). Since the core content is identical in the two versions of each example, these changes cannot be attributed to a difference in difficulty between the groups.

For ResNet-50 on Waterbirds, the Figure plots, for each test bird, its margin on the shortcut-aligned background against its margin on the conflicting background. A model that ignores the background would lie on the diagonal: the ERM points lie clearly below it, whereas the GAP-FSO points lie close to it. Since GAP-FSO also lowers the overall margin scale (the mean absolute margin falls from 8.05 to 3.42), we also report scale-free measures: the sensitivity relative to the mean absolute margin falls from 0.455 to 0.224, and the flip rate falls from 11.9% to 4.4% (Table 5). The reduction holds in both classes (cf_gap of landbirds 3.48 to 0.73, waterbirds 4.26 to 0.87), which means that the prediction depends less on the background and more on the bird.

To evaluate whether reductions in the gap Δ\Delta, i.e., the difference in the average signed margin yzyz between the aligned and conflicting groups, actually reflect reduced reliance on shortcuts, we compute the Pearson correlation between the training gap and the counterfactual sensitivity across all GAP-FSO runs (spanning all values of γ\gamma and random seeds). Specifically, each run checkpoint yields a pair of measurements: (1) the gap Δ\Delta evaluated on the training set, and (2) the counterfactual sensitivity of the checkpoint (cf_gap for the image tasks and cf_abs for CivilComments). As shown in Table 8, this correlation is strong on most image datasets, for example Colored-MNIST (0.910), Dogs-vs-Cats (0.639 and 0.783 for ResNet-50 and CLIP ViT-B/32), and Waterbirds (0.720). However, it is weak on CivilComments (0.224 to 0.305). A plausible reason is that identity terms in text are harder to isolate as a single shortcut than visual attributes. Because Δ\Delta tracks the counterfactual behavior on most image datasets but only weakly on text, we rely directly on the counterfactual outcomes rather than on the magnitude of Δ\Delta alone as evidence of shortcut mitigation. We have incorporated these correlation results and analysis into the revised manuscript.

Table 8: Pearson correlation between the gap Δ\Delta computed on the training data of the checkpoint and the counterfactual sensitivity (CE loss).

Model and datasetSensitivityPearson correlation
CNN, Colored-MNISTcf_gap0.910
ResNet-50, Dogs-vs-Catscf_gap0.639
CLIP ViT-B/32, Dogs-vs-Catscf_gap0.783
CLIP ViT-B/32, Waterbirdscf_gap0.720
DistilBERT, CivilCommentscf_abs0.305
Qwen2.5-0.5B, CivilCommentscf_abs0.224
AuthorsOpenReview ↗

References

[1] Hermann, K., Mobahi, H., Fel, T., & Mozer, M. (2024). On the foundations of shortcut learning. In International Conference on Learning Representations (Vol. 2024, pp. 43832–43868).

[2] Qiu, G., Kuang, D., & Goel, S. (2024). Complexity matters: Dynamics of feature learning in the presence of spurious correlations. arXiv preprint arXiv:2403.03375.

[3] Lim, J., Kim, S., & Moon, S. M. (2026). Shortcut Features as Top Eigenfunctions of NTK: A Linear Neural Network Case and More. Advances in Neural Information Processing Systems, 38, 94131–94159.

[4] Park, S., Kim, J., Lee, H., & Song, K. (2026). Spurious correlation-aware embedding regularization for worst-group robustness. In International Conference on Learning Representations (Vol. 2026, pp. 156420–156445).

[5] Gretton, A., Borgwardt, K. M., Rasch, M. J., Schölkopf, B., & Smola, A. (2012). A kernel two-sample test. The Journal of Machine Learning Research, 13, 723–773.

AuthorsOpenReview ↗

Appreciation and Closing Remarks

We have incorporated all experiments and analyses from this rebuttal into the revised manuscript (marked in blue). Because the rebuttal period is short, we could not complete all experiments immediately; we will do our best to finish the remaining ones and include them in the final version. We have done our best to explain our work, and we hope that the experiments and explanations above help you better understand this study and its contribution. We also welcome further discussion, will answer any question seriously, and would be grateful if you would reconsider your evaluation in light of these responses. Thank you!

Thank you for the additional counterfactual experiments. I find the evidence that GAP-FSO reduces reliance on certain shortcuts more convincing, but when the groups differ in difficulty, the gap is still an unreliable measure of shortcut strength. The experiments also do not directly show that GAP-FSO alleviates causal starvation during training, so I still have reservations about the proposed mechanism and will keep my original rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.